An Autonomous Vehicle Prototype for Off-Road Applications based on Deep Convolutional Neural Network

Abstract

Making autonomous driving a safe, feasible, and better alternative is one of the core problems for researchers from academia and industry. The development of autonomous cars in a real-world setting poses many challenges. This paper presents a low-cost, small-scale self-driving automobile model called ‘Zumee’, powered by a deep convolutional neural network. Zumee offers synchronised image data collection using a camera for deep learning models. The proposed model is experimentally validated using an Nvidia Jetson TX2 board as the onboard computer. Data augmentation, transfer learning, and neural networks are used to train the prototype model. Specifically, a convolutional neural network (CNN) model is trained using data from various scenarios. A robust dataset is generated by augmenting the images obtained using a camera to accommodate multiple environments. The prototype model is trained to imitate the policies used by a human supervisor to drive a robot car autonomously in a closed environment. The proposed prototype could be used as a base for more advanced driverless autonomous vehicles or for educational purposes as an economical prototype model.